In this article
For a busy professional, this can be a modest exercise. You do not need to launch a company-wide innovation programme to reconsider an awkward customer handover or an ineffective team briefing.
Give AI for innovation a clear problem boundary
Describe the friction, the people affected and the outcome you want. Keep the proposed solution out of the first question where possible.
In a hypothetical service team, “How can we create a better dashboard?” presumes that a dashboard is needed. “How could a new shift understand unresolved customer issues without reading every message?” leaves room for several approaches.
Ethan Mollick's innovation-through-prompting essay explores AI-supported idea generation. The practical method here is an original workplace exercise: separate problem definition, option generation and testing.
Write down the constraints before requesting ideas. Examples include no new budget, no customer information in unapproved tools and no extra meeting time.
Generate differences, not just a longer list
Ask for options that vary in mechanism. Ten versions of “improve communication” are not ten meaningfully different possibilities.
Try this prompt using fictional or approved information:
Suggest distinct ways to help an incoming shift understand unresolved work. Group the ideas by mechanism: changing the information, changing ownership, changing timing and removing unnecessary work. For each idea, state an assumption and a reason it might fail. Respect the constraints below.
Review whether the groups truly differ. A new template and a new colour scheme may be the same underlying idea; a designated handover owner changes the process more substantially.
You can also reverse the assumption: “What would we try if we could not add another form?” This can expose alternatives that a tool-led discussion overlooks.
Make constraints productive
Use constraints to sharpen an idea rather than asking for an unrealistic ideal. The aim is a workable possibility in your setting.
For the hypothetical shift team, an elaborate app may be impossible to maintain. A short unresolved-items view within an existing approved system may be easier to test. Neither is automatically the right answer; the available infrastructure and people's needs determine that.
Ask the tool to identify the smallest change that would test the important assumption. Avoid letting it turn every idea into a large transformation programme.
Evaluate with people who know the work
Select a few criteria before choosing a favourite. Consider usefulness, feasibility, burden on colleagues, information risk and ease of reversing the change.
| Question | What it protects against |
|---|---|
| Who benefits, specifically? | An attractive idea with no clear user |
| What must be true for it to help? | An untested assumption |
| Who does extra work? | Hidden effort transferred to someone else |
| How could we test it safely? | Premature commitment |
Ask a colleague who receives the output to challenge your preferred idea. Their objection may reveal a requirement that neither you nor the model considered.
Do not treat an AI-generated market forecast, customer quote or invented case study as validation. Get evidence through an appropriate real-world test.
Design an experiment around one assumption
Make the test smaller than the proposed rollout. Decide what you will observe, how long you will run it and what would make you stop or revise it.
In the hypothetical handover example, the assumption might be that separating unresolved items from routine updates helps the next shift find important work. A bounded pilot could compare how colleagues locate those items using approved synthetic scenarios before any operational change.
Record both the useful result and the additional effort. A format that improves clarity but demands unrealistic preparation may need redesign.
NIST's AI Risk Management Framework is a reference for risk-aware AI use. Your experiment still needs task-specific limits and accountable people.
Turn one idea into a test card
Write a small card with five fields: idea, assumption, test, observation and decision. Keep the observation separate from your preferred interpretation.
For the fictional shift handover, the idea is a dedicated unresolved-items view. The assumption is that colleagues miss important work because it is buried among routine updates. The test uses several synthetic handovers, with participants explaining what they would act on next.
Observe whether they find the unresolved work, whether ownership is clear and what preparation the format requires. Do not manufacture numerical improvement claims from a tiny informal exercise. A useful observation can be qualitative and specific: one field was repeatedly misunderstood, or a missing owner became easier to notice.
Then choose the next decision. You might revise a field, test with another role or abandon the idea because the real obstacle is unavailable information.
This card keeps the generated proposal attached to a question you can answer. Without it, an interesting idea can survive indefinitely as a persuasive slide, gathering enthusiasm without accumulating evidence.
Keep ownership of the final choice
AI can help generate objections and alternative designs, but it should not decide whose workload or customer experience changes. Explain the selected option in terms the affected people can question.
Choose one recurring frustration and ask for options that change different parts of the process. Take only one promising assumption into a bounded test.
For further reading, explore PiOfAI, building a personal AI workflow and leading change without losing people. Useful innovation ends in evidence about a real problem, not a longer idea list.
AI-assisted researched guidance from PeakBelt. Hypothetical examples are labelled in the text. See our editorial approach.